Two questions come up in every AI conversation with a Malaysian IT leader. What does it cost, and where does the data go? It’s actually quite straightforward: a public model gets a proof-of-concept running in an afternoon, while a private one keeps sensitive data inside the building. Both answers are correct, just being designed for different workloads.
Before committing to one, or if you should just solely pick one, it helps to understand the nuances in private AI versus public AI, when a hybrid is the honest answer, and how business automation and AI solutions in Malaysia can be structured around the split.
Public AI, in Practical Terms
Public AI means using a hosted model through an API. OpenAI, Anthropic, Google, and Azure OpenAI are the common ones. Billing is per token, so cost scales with how much you use it. As the model is hosted by a vendor, setup is fast and you do not manage the model, the GPUs, or the updates.
The trade-off is that prompts and responses leave the environment. While vendors do publish enterprise data-handling policies, and the major providers offer zero-retention options, the data path still crosses a third-party boundary.
For low-sensitivity work like summarising public documents or drafting marketing copy, that is usually fine. But for sensitive data like customer records, medical notes, or financial statements, maintaining compliance in privacy and data protection becomes the most important aspect.
Private and Self-Hosted AI, in Practical Terms
A self-hosted LLM runs on infrastructure the business controls. That can be a server in the office, a Malaysian data centre, or a private cloud tenancy. With this setup, data stays inside the boundary, and costs shift from per-token to per-hour of GPU capacity plus maintenance.
Free open-source models made this practical. DeepSeek, Meta’s Llama, and Alibaba’s Qwen can all be downloaded and run in-house at no licence cost. The smaller versions run on a single business-grade GPU and handle most everyday tasks. The largest versions match the top paid models, but they need much more hardware to run well, and that is where the ongoing cost sits.
With that, the real cost of a local LLM or on-premise AI deployment is not the model itself, but the GPU capacity to run it at business hours peak, the monitoring to keep it up, and the person or team who owns the updates.
Why Malaysian Firms Land on a Hybrid
For an IT director, the ideal setup usually encompasses both a public and a self-hosted model. A local model handles the high-volume, low-complexity work where data sensitivity matters. On the other hand, a public API handles the low-volume, high-complexity work where the reasoning has to be top-tier and the data can be sanitised first.
A specific example: a logistics firm runs invoice extraction on a self-hosted model, because thousands of invoices a day through a public API is neither cost-effective nor compliant. The same firm uses a public API for the occasional contract analysis, because the reasoning is more demanding and the document is already reviewed for what leaves the building. Two models, two use cases, one architecture.
The market signal is consistent. AWS research from late 2025 found 67 per cent of Malaysian businesses that adopted AI expect an average of 15 per cent in cost savings. That number holds only if the split between paid and self-hosted is the right one for the workload.
Where PDPA and RMiT Change the Answer
The Personal Data Protection Act 2024 amendments have been in full effect since 1 June 2025. Cross-border transfer rules are stricter, and the 72-hour breach notification clock starts at awareness. Anything an AI model reads, writes, or infers about a person is in scope.
For financial institutions, Bank Negara’s Risk Management in Technology framework adds a layer. Outsourcing to a public API means treating the vendor as a third-party service provider, with the assessments and controls that go with it. For sectors under the Cybersecurity Act 2024, similar residency and reporting logic applies.
None of this rules out using a public API, but it does set the line for data classification, and what needs to stay behind it.
A Simple Decision Framework
The way to cut through the vendor pitch is to answer four questions for each use case.
- Data sensitivity. Would a leak trigger a PDPA notification or breach a client contract? If yes, default to local.
- Volume. Is the workload thousands of calls a day, or dozens? Volume favours self-hosted; low volume favours per-token billing.
- Reasoning depth. Does the task need frontier-model quality, or is a distilled open model enough? Most business tasks fall into the second bucket.
- Latency and uptime. Does the workflow break if the model is down for an hour? Local gives more control; hosted spreads the risk.
If two decisions lean local, self-hosting is usually the better route. Likewise, two toward public usually means an API. A mix means hybrid, which is where most production estates end up.
Getting the Split Right
The wrong architecture shows up quickly. A private model handling low-volume marketing tasks wastes GPU capacity, and a public API handling every customer record is a compliance issue waiting to be discovered. The right split maps each workload to the model that fits it, then keeps them in one operational stack.
Cue Net Onboard, where our AmplifyChampion team scopes which workloads sit behind a private model, hosts the model on the right infrastructure, and manages the public API layer for the workloads that need it. Along with our AmplifyControl pillar, we maintain data protection that covers what crosses the boundary and what stays inside.
Get in touch with our team or WhatsApp us about business automation and AI solutions in Malaysia to work out where the split between public and private AI should sit for your specific business.